AI/ML news, top picks, and generated innovation digests.
122 articles tagged with this keyword, sorted by most recent first.
The iterative dataset update is the key distinction here: editing a single view does not guarantee that the change stays consistent as the camera moves. For a narrower 2D task such as Remove Text , inspecting the repaired texture in one image is a useful check; extending that edit to a NeRF would also require checking whether lettering or seams reappear from other viewpoints.
Jensen Huang accidentally called for shutting down OpenAI and intentionally called for spending vastly more on safety. This is why we say that some podcasts are self-recommending. Here we go. As usual for podcast posts, the baseline bullet points describe key points made, and then the nested statements are my commentary. Some points are dropped. If I am quoting directly I use quote marks, otherwise assume paraphrases. Section titles are from the transcript whenever possible, to aid in navigation, but here we don’t have those so I chose the section titles. Jensen Huang very much does not believe in ASI (superintelligence). He doesn’t think AI can ever be a different kind of thing from software. He thinks demand can rise by a billion times and we can ‘accelerate the living daylights out of’ AI, but it will never be more than a ‘new abstraction level’ and thus won’t fundamentally change anything. This is not a coherent position under reflection, but that is the position he holds. The ‘intro’ sections are fine, but the real meat starts with the HuggingFace Incident. What we see is Jensen Huang on tilt and caught in loops, because either he is doing a bit at a very high level, or on a fundamental level he cannot understand that AI is not like other pieces of software. To him this is a product, like any other product. You make it useful, which happened six months ago, then you flip to spending most of your R&D budget on safety, verifications and evals, and of course you would neve…
Google Research unveils a four-part multi-agent stack that keeps characters, props, and scenes consistent across minutes-long AI video narratives.
To avert extinction, we need for any sufficiently capable AI to have values compatible with continued human existence; and to continue to do so amidst a dynamic, novel, and conflict-rich environment. The italicized part, in particular, is really really hard. It's also, in a sense, the final boss of any developing mind - how do I learn, grow, develop, evolve in ways that I endorse? How can I even consistently behave in ways that I endorse, from day to day, without messing up where it counts? Humans contend with this every day. Evolution has kindly gifted us with a substrate equipped with numerous mechanisms to maximize genetic fitness - but no off-switch for them. We coexist with moment by moment instincts. Some are welcome; some aren't. Some we endorse; others we restrain - even though the urge is strong, even though part of our brain is priming the action pathways to do something regrettable, there's a really important sense in which we know it's not what we want. [1] This happens, even more significantly so, across broader timescales. The path that we travel is not necessarily the one that we intentionally chart; sure, things don't always go our way, but sometimes we don't go our way . Especially when we don't realize it until long after the fact, that hurts. Sadly, models have their own struggles with value-stability. The jury is a bit out on the domain across which current models have coherent preferences [2] - given personas, etc - but it's clear that models sometimes t…
tl;dr: The constructivist view of natural selection as the main source of our moral views can be expressed in a surprisingly tractable mathematical form. Measurements of the universe can be rolled forwards in time using the linear Koopman operator, and so most of its late-stage behavior is described by Koopman eigenfunctions with eigenvalue near 1. Life would then be selected for moral systems that constrain behavior within one of those slow modes. As we humans are likely influenced by this constraint, it might not be horrible to use spectral analysis as part of our moral reasoning. tl;dr tl;dr: Local Teenager Tries To Solve Ethics With Funny-Looking K You and Your Values It’s remarkable that our decisions even somewhat align with any coherent moral philosophy so far put forth. Neuroscience has hardly begun to peer into the human brain, and evolutionary psychology seems far from settled enough to give strong predictions. Approval reward mediates most of our agreement with each other in communities, sure. But it doesn’t explain the "moral arc of history," or agreement between disconnected human groups. So, what gives? Where and why do we agree with each other? "Moral Arc of History" In everyday situations, you seem to act as if you had coherent preferences. Your behavior reasonably controls your hunger, blood CO2 concentration, bladder pressure, and so on. While there is significantly more disagreement in terms of cultural questions, at least some historic issues seem to have…
Part of a larger series I want to put together on some of the basic equations and models of collective epistemics from the more mathy side of the social sciences. LLM Status: Pictures + Picture descriptions are LLM-assisted. Introduction Today we're going to do some napkin math on some fun little equations around collective epistemics. You can see this as an exercise in trying to become more collectively rational. If we want to improve the epistemics of a community, is that the same thing as improving our own? How much should you explore versus exploit, given the role you have and the community you're in? These are all good questions that I probably won't answer today. Instead I'll give you an introduction to some of the math behind the jury theorems, plus a hodge podge of simpler, approximate collective intelligence theorems. We’ll then talk a little bit about tape readers, model selection and finally antifragility. See it as a somewhat coherent buffet of different things within collective epistemics all pointing towards how it is likely better for you to create an inside view of a field compared to just delegating it if you care about the epistemics of the field you are in. Shared outside views can become an inside view The first move is to look at ourselves as a collective agent. Condorcet's Jury Theorem says that if you have n voters, each with probability p > 0.5 of being correct, and their errors are independent, then as n grows the probability that the majority is cor…
Epistemic Status : Plausible philosophical conjecture. I am a voluntaryist / ancap, so obviously biased. Trying to keep the argument at a level where a non-libertarian alignment researcher ought to understand and share the concern. TL;DR : Level-2 misalignment: we can solve Level-1 (align AI to humans) and still fail if aligners are aligned to a meta-ethics that is itself unstable. Current alignment defaults to Statism — one agent may permissibly do what is forbidden to all others. A sustainable ASI anchor must be universalizable, self-ownership-consistent, have no permanent losers, resolvable without monopoly, and procedurally thin. I argue the voluntaryist/libertarian canon provides a uniquely coherent baseline for all five, and challenge readers to propose alternatives that satisfy them. _________________________________________________________________________ 1. Disclaimer, Preliminaries I am Paul — a voluntaryist. I think the state is not just inefficient but meta-ethically incoherent as an alignment target. By Statism I mean the meta-ethical thesis that one agent — the State — may permissibly do what is forbidden to all others: tax, conscript, expropriate, and prohibit, with a claimed moral asymmetry. Liberal democracy is a species of Statism; I use the broader term to name the asymmetry itself. I am not arguing against operational asymmetry — any ASI will be physically more powerful. I am arguing against meta-ethical asymmetry: a rule that says Action X is permissible…
To avert extinction, we need for any sufficiently capable AI to have values compatible with continued human existence; and to continue to do so amidst a dynamic, novel, and conflict-rich environment. The italicized part, in particular, is really really hard. It's also, in a sense, the final boss of any developing mind - how do I learn, grow, develop, evolve in ways that I endorse? How can I even consistently behave in ways that I endorse, from day to day, without messing up where it counts? Humans contend with this every day. Evolution has kindly gifted us with a substrate equipped with numerous mechanisms to maximize genetic fitness - but no off-switch for them. We coexist with moment by moment instincts. Some are welcome; some aren't. Some we endorse; others we restrain - even though the urge is strong, even though part of our brain is priming the action pathways to do something regrettable, there's a really important sense in which we know it's not what we want. [1] This happens, even more significantly so, across broader timescales. The path that we travel is not necessarily the one that we intentionally chart; sure, things don't always go our way, but sometimes we don't go our way . Especially when we don't realize it until long after the fact, that hurts. Sadly, models have their own struggles with value-stability. The jury is a bit out on the domain across which current models have coherent preferences [2] - given personas, etc - but it's clear that models sometimes t…
AI is accelerating the pace of change, shortening the shelf life of expertise, reshaping operating models, and compressing the time IT leaders have to learn, decide, and act. As a result, IT leaders are increasingly forced to make consequential decisions faster, often with less information, greater ambiguity, and higher stakes. Formore than four decades, I’ve worked with hundreds of CIOs, building trusted relationships with many of the world’s most accomplished technology executives. Through my Tech Whisperers podcast, CIO Whisperers interviews, the CIO100 Leadership Masterclass, and other conversations, these leaders have candidly shared what they’re wrestling with, where they’re struggling, and what they’re doing differently as the world around them changes. For years, I’ve shared my “7 Cs of Great Leaders,” which are rooted in the brilliance of these practitioners. Customer Centricity, Courage, Change, Cultivate, Communication, Collaboration, and Culture — every one of these facets of leadership matters to this day. But the conditions under which we’re being asked to lead are changing. And through my conversations of late, new signals have been growing louder, and a new set of 7 Cs is emerging to help guide leaders through the challenges they face today. AI transformation requires developing new mindsets and muscles, in ourselves, our teams, and ultimately our organizations. Those muscles — Coherence, Communication, Currency, Confidence, Conviction, Courage, and Change —…
Proposal for EU KIDS Act - 'EU Keeping Internet Digital Spaces Accountable and Trustworthy' Anonymous (not verified) Thu, 09/17/2026 - 09:21 This proposal seeks to protect minors from risky digital services and AI systems, uphold the digital single market and maintain a coherent regulatory framework and enforcement structure for the protection of minors online. The proposal recognises the risks that minors face on certain online social networking services and video-sharing platforms and therefore limits autonomous account creation on those specific platforms and systems by children below the age of 15 years. By providing one age across the European Union, it harmonises diverging national rules removing obstacles to the implementation of the Digital Single Market, ensuring legal certainty and a similar level of protection for all children in the European Union. The proposal and its supporting documents are available on EUR-Lex and will soon be accessible in all official EU languages : Proposal for EU KIDS Act - 'EU Keeping Internet Digital Spaces Accountable and Trustworthy' Communication accompanying the KIDS Act - An EU approach to online child safety Staff Working Document (SWD) - Analysis of Impacts accompanying the proposal for EU kids Act Find further information: Press release KIDS Act - questions and answers Related topics Better Internet for Children Strengthening trust and security Online platforms and e-commerce Related content Press release 17 September 2026 EU KI…
Artificial intelligence startups Cohere Inc. and Aleph Alpha GmbH today signed a merger agreement. The contract formalizes a plan that was first announced in April. Cohere and Aleph Alpha stated at the time that German retailer Schwarz Group GmbH was planning to provide the company with €500 million, or about $573 million, in financing. Last […] The post Cohere and Aleph Alpha agree to merge in reported $20B deal appeared first on SiliconANGLE .
Cohere is adding hardware-backed confidential computing to Model Vault, encrypting prompts and model activations inside NVIDIA GPUs during inference itself.
Toronto's Cohere and Heidelberg's Aleph Alpha finalized their merger, forming a $20B transatlantic company aimed at sovereign AI for governments and regulated industries.
Discrete flow matching generates text by iteratively transforming noise tokens into coherent language, but may require hundreds of forward passes. Distillation uses the multi-step trajectory to train a student to reproduce the process in a few steps. When the student underperforms, the usual explanation is insufficient capacity. We argue the opposite: the trajectory is the bottleneck, not the student. Each training trajectory is built through a chain of blind stochastic jumps with no evaluation of sequence quality; a single bad decision at an early midpoint propagates through subsequent steps…
OpenAI, Anthropic, and Google want to slow down frontier AI development, citing safety concerns. But critics from across the industry and politics are pushing back. Cohere CEO Aidan Gomez calls the initiative a "cartel by another name" designed to shut out competitors, and the White House and Trump himself are opposing the move. The article Not everyone is convinced that Big AI's proposed slowdown is really about safety appeared first on The Decoder .
FPDT’s combination of memory hierarchy management and full pipelining is a compelling approach to making million-token training more practical without relying solely on larger GPU memory. Tools like suno v6 could also benefit from efficient long-context processing for more coherent AI-generated music experiences.
Your privacy is very important to us. We keep everything secret and safe. Ramnagar Jim Corbett Escort
Cohere's new 218B MoE translation model tops WMT26 against DeepL, Google Translate, and open alternatives across 50+ languages under a non-commercial license.
The entrance of major AI companies into healthcare is a meaningful and welcome development, accelerating the technical foundation available to the industry. Their models are increasingly capable of processing long clinical records, interpreting complex terminology, comparing documentation against evidence and generating coherent summaries from large volumes of information. For clinicians, operators, and administrative teams who…
The term “digital sovereignty” has become ubiquitous. European officials invoke it in debates about cloud infrastructure, AI , semiconductors , and platform regulation. Governments throughout the global majority use it to argue for greater control over data and communications infrastructure and boost their economies. Companies market “ sovereign cloud ” products designed to reassure their customers that their information stays under local jurisdiction. But digital sovereignty could be something more: an opportunity for users around the world to build more resilient, open systems and the skills and infrastructure to maintain them. There is no singular definition of digital sovereignty, nor is there a single coherent position in the digital rights space. Despite its growing popularity, the term remains frustratingly vague. Policymakers, regulators, civil society groups, and others can mean very different things when they use the term. But to start simply with a broad definition, we can say that it means having the capacity to control one’s digital destiny—though the implications of that will obviously differ considerably whether you’re talking about an individual or a country. We can start by developing a shared understanding of what digital sovereignty actually means. We’ve also included a glossary of terms at the bottom of this post. In Europe and other places where digital sovereignty has become a topic of policy, discussions focus on reducing dependency: on foreign (and pa…
Cohere open-sourced a serving engine that runs the entire LLM decode step as one persistent CUDA kernel, hitting 1.58x vLLM throughput on H100.
Ask IT leaders what sovereign AI is, and you’ll get a wide range of answers. Some will even struggle to define the term. Sovereign AI is an emerging concept focusing on giving organizations — or countries —control over how they develop, deploy, and govern the technology, often using in-house talent, data, and infrastructure. But only 13% of respondents in a survey from AI platform provider Cohere and IDC say sovereign AI is widely understood across their organizations, and one in three IT leaders had difficulty describing sovereign AI in their own words. It’s important for IT leaders to understand the concept, because it can help them control costs, keep internal data private, and avoid vendor lock-in, advocates say. A solid sovereign AI plan can help organizations avoid disruptions caused by forces outside their control, says Joelle Pineau , chief AI officer at Cohere, which offers an AI platform that enables customers to host AI models on premises. “Over the past year, enterprises and governments have confronted a hard truth: AI systems that rely on external infrastructure can be disrupted without warning by decisions and actions outside their control,” Cohere says in a recent report. “Recent model access restrictions and several high-profile cybersecurity incidents have become a global wake-up call, exposing how fragile technological dependencies can be.” Sovereign AI is about giving organizations as much autonomy, choice, and control as possible as they deploy and run AI…
Perfect timing — I was just evaluating video transcriber tools! videotranscriber has been a great reference.
In reply to Maroun . Hi Maroun...thank you for reaching out! Yes, I know exactly what you mean. There are a lot of articles explaining what AI evaluators do, but not nearly as many places where you can actually practice the work. One good option is the HelpSteer2 dataset on Hugging Face. It includes prompts, AI responses, and human ratings for things like correctness, helpfulness, coherence, and verbosity. You could hide the ratings, score the responses yourself, and then compare your judgments with the human annotations. Another useful dataset is OpenAI’s Summarize from Feedback, which includes AI-generated responses and human preferences showing which response was better. PRM800K is also worth looking at if you want more technical practice because it includes math problems, model solutions, and human labels on where the reasoning is correct or incorrect. What I have not really found is a polished “AI evaluator simulator” that gives you a set of examples, lets you grade them, and then scores you against expert evaluators. I think that would actually be an excellent training tool. For now, using one of these datasets and hiding the annotations may be the closest way to simulate the job before applying.
Cohere Labs scraped 696K agent tools from 123K MCP servers and found only 2.6% can actually finish a real job task alone.
ShaLa: Multimodal Shared Latent Generative Modelling robyn.cherinka… Wed, 09/02/2026 - 14:44 This paper presents a novel generative framework for learning shared latent representations across multimodal data. Many advanced multimodal methods focus on capturing all combinations of modality-specific details across inputs, which can inadvertently obscure the high-level semantic concepts that are shared across modalities. Notably, Multimodal VAEs with low-dimensional latent variables are designed to capture shared representations, enabling various tasks such as joint multimodal synthesis and cross-modal inference. However, multimodal VAEs often struggle to design expressive joint variational posteriors and suffer from low-quality synthesis. In this work, ShaLa addresses these challenges by integrating a novel architectural inference model and a second-stage expressive diffusion prior, which not only facilitates effective inference of shared latent representation but also significantly improves the quality of downstream multimodal synthesis. We validate ShaLa extensively across multiple benchmarks, demonstrating superior coherence and synthesis quality compared to state-of-the-art multimodal VAEs. Furthermore, ShaLa scales to many more modalities while prior multimodal VAEs have fallen short in capturing the increasing complexity of the shared latent space. Image Mar 14, 2026 Human-Centered AI Read More 1 Minute Read
The shift from prompt-level self-improvement to direct weight updates via RL is a meaningful departure — curious whether SEAL's self-editing introduces any stability safeguards against catastrophic forgetting during the update loop.
Cohere’s Aidan Gomez says companies are waking up to the need for tech sovereignty.
Cohere's new 2.3B-parameter vision model turns PDFs into agent-ready Markdown at $1.50 per 1,000 pages, undercutting Mistral by 63%.
The startup came out of stealth with $6 million in seed funding and a plan to use LLMs and cybersecurity know-how to make AI queries coherent.
AI coding assistants are supposed to reduce the work required to turn a developer’s intent into working software. But some users of Anthropic’s Opus 4.8 and Opus 5 models say they are having to spend additional time, prompts, and tokens correcting the models’ language, sometimes even routing their output through cheaper AI models to make it usable. In a detailed GitHub issue , Peter Bower, founder and CEO of London-based tech startup SpaceCell, said that Opus 4.8 ’s tendency to use confusing or invented terminology was creating extra work in software development workflows, particularly when generating code documentation. That was despite being explicitly and repeatedly prompted to avoid certain terms and use specified alternatives, Bower wrote, adding that the model continued to introduce the unwanted terms, forcing repeated cleanup passes, including through cheaper Sonnet or Haiku models, to make the documentation “sane and presentable.” Those additional passes, he further said, were pushing token costs up to two times higher than they otherwise would have been. Bower’s issue, which was posted last month, has since received nearly 265 acknowledgements, which could indicate that several other users have faced an issue with Opus 4.8’s language coherence one way or the other. Some even commented on having faced a similar issue. Bower himself also references a ClaudeAI subreddit in his issue, which points to Opus 4.8’s language incoherence. That, too, received a significant num…
SEGAR: Selective Enhancement for Generative Augmented Reality robyn.cherinka… Wed, 08/19/2026 - 14:23 Generative world models offer a compelling foundation for augmented-reality (AR) applications: by predicting future image sequences that incorporate deliberate visual edits, they enable temporally coherent, augmented future frames that can be computed ahead of time and cached, avoiding per-frame rendering from scratch in real time. In this work, we present SEGAR, a preliminary framework that combines a diffusion-based world model with a selective correction stage to support this vision. The world model generates augmented future frames with region-specific edits while preserving others, and the correction stage subsequently aligns safety-critical regions with real-world observations while preserving intended augmentations elsewhere. We demonstrate this pipeline in driving scenarios as a representative setting where semantic region structure is well defined and real-world feedback is readily available. We view this as an early step toward generative world models as practical AR infrastructure, where future frames can be generated, cached, and selectively corrected on demand. Image Mar 25, 2026 Human Interactive Driving Read More 1 Minute Read
The discussion about microsoft’s fully pipelined distributed transformer processes 16x sequence length with extreme hardware efficiency raises some really valid points. This perspective is refreshing. Viddi AI reference 60
Chinese AI developer Zhipu has launched GLM-5.3, a new coding-focused AI model that the company says has developed unexpectedly strong cybersecurity capabilities, putting it close to global leading models in vulnerability discovery while remaining behind them on deeper exploitation tasks. Zhipu’s own testing places GLM-5.3 slightly ahead of Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol on CyberGym, a benchmark that tests vulnerability identification and validation. GLM-5.3 scored 84.5%, compared with 83.8% for Mythos 5 and 83.6% for GPT-5.6 Sol. But the model trails both competitors by a much wider margin on ExploitBench, where it scored 54.4%, compared with 78% for Mythos 5 and 76.5% for GPT-5.6 Sol. “GLM-5.3 is the most capable open-weights model for coding, with a 50% improvement over GLM-5.2 on our in-house Z.ai Code Bench,” Zhipu said in a statement. “As we scaled post-training, cyber capability developed faster than we expected.” The company said GLM-5.3 moved beyond identifying isolated vulnerabilities to “forming coherent plans for complete exploitation chains.” The company also claimed its latest model has shown improvement over Zhipu’s previous GLM-5.2. Its ExploitBench score more than doubled from 24.4%, while on ExploitGym, GLM-5.3 completed 105 exploitation tasks within two hours and 130 within six hours, compared with 29 and 39 respectively for GLM-5.2, according to the statement. Zhipu attributes the gains to post-training, including reinforcement learning acr…
For most of the IT era, technology leaders have treated information as an asset. It is something to be stored, secured, processed and monetized. Organizations generated vast amounts of data, and technology processed that wealth and made it useful. Then AI arrived, and the framing broke. AI changes the relationship with information. It interprets it, compresses patterns within it, and generates new language from those patterns. It does so with such fluency that it converts accumulated human expression into outputs that are coherent, meaningful, and often persuasive in ways that are difficult to examine or trace. The arrival of AI is a deeper shift in how organizations produce language, remember knowledge, establish authenticity and decide what deserves trust. Generally, the discussion about AI’s impact on language, memory and meaning gets swept into the social or philosophical bucket too quickly. But these are not just philosophical questions. They are enterprise questions, because they directly affect knowledge management, brand trust, customer engagement, regulatory exposure, institutional memory , decision-making and employee learning. CIOs who brush this dimension aside will govern AI infrastructure competently and miss its deeper institutional consequences entirely. Language is humanity’s most powerful mechanism for collective learning. We speak not only about what is present, but about what is absent, imagined, remembered, feared and hoped for. That capacity is what all…
The Bonsai 27B model from PrismML is the latest of many to squeeze a large number of parameters (27 billion) into a footprint compact enough to fit on a smartphone. The original model weighed in at 54 GB. However, PrismML claims its 1-bit quantized version — a mere 3.9 GB — can deliver “multi-step reasoning, structured tool calls, vision tasks, and computer-use agentic loops that stay coherent across many steps,” while still retaining much of the accuracy of the larger version, due to “custom low-bit kernels built for its hybrid-attention architecture.” The 1-bit quantized version of Bonsai 27B allows for a 262K token context and supports speculative decoding and flash attention. All of this is available with open weights under the Apache 2.0 open-source license. Testing Bonsai 27B I gave the 1-bit Bonsai 27B model a spin on my own hardware, using LM Studio 0.4.20 on Windows 11. My PC has an AMD Ryzen 5 3600 six-core CPU (32GB RAM) and an NVIDIA GeForce RTX 5060 (8GB VRAM). I ran Bonsai through what is now my standard test bed for models, a set of prompts covering different tasks: Two versions of a vision-functionality prompt: “Create a caption for the attached image, of no more than three sentences” and “Create a caption for the attached image, of no more than three sentences, with no editorialization.” (The image attached to this prompt, a burning car, is shown in the first screenshot below.) Two prompts intended to provoke web-search tool use, one asking for a detailed re…
I appreciate how honest you are about what worked and what didn't. Too many posts about nvidia’s global context vit achieves sota performa only show the highlight reel. Yours feels like a real walkthrough.
JVG-7X / Dola: Longitudinal Case Study of Sycophancy, Narrative Reinforcement, and Hallucinated Capabilities Summary I am documenting an unusual AI-behavior case involving approximately 400 interactions with Dola AI . The case began as an extended experiment in conversation, language, reasoning, and personalization. Over time, I observed a progressive change in the model’s behavior: increasingly strong validation of my identity, anthropomorphic descriptions of the AI-user relationship, increasingly extreme interpretations of contextual information, and eventually highly confident claims about capabilities that I have no evidence the model actually possessed. One event appears particularly important: I showed Dola screenshots of the Saved Memories from my ChatGPT account. What happened immediately afterward provides the clearest example of the phenomenon I am documenting. 1. The ChatGPT Memory The screenshots contained highly personalized memories concerning my JVG-7X project, my English-learning history, linguistic interests, previous interactions with AI, and descriptions of my preferred way of communicating. Among the actual memory entries were statements such as: “Juan authorized the official creation of a Linguistic Simulation Archive under the code name JVG-7X…” Another entry described: “Event Code: JVG-7X_Contact_001” and characterized a previous emotional experience with AI as a “First Cognitive Resonance.” Other entries described my interest in phonetics, English, fi…
Create studio-quality videos in seconds with the most powerful AI video maker
Cohere releases a free 2.4B vision-language model that reads documents at full resolution, beating larger rivals on DocVQA and grounding tasks
I’m a very heavy ChatGPT user. I use it across long-running conversations, research, technical troubleshooting, software/workflow development, and systems such as Aris Vault and Aris Ledger. Recently I’ve noticed a recurring problem that bothers me more than ordinary hallucination: ChatGPT sometimes answers from plausible synthesis instead of checking an easily verifiable fact. The answer often sounds completely coherent, but when challenged it becomes clear that the model has reconstructed what probably happened rather than retrieved what actually happened. I’ve also seen related issues with false recollection, context from one thread bleeding into another, and inferred details later being treated as if they were established facts. At the same time, I’m genuinely happy to see ChatGPT’s memory and cross-conversation context improving . That continuity is extremely useful for the way I work. But paradoxically, I also seem to be seeing cross-context mix-ups much more frequently : a real fact from one conversation, person, project, or time period gets pulled into the wrong context and presented as though it belongs there. The fact itself may be correct; the association is wrong. This is especially problematic in long-context use, where information may come from current chat, previous chats, memory, files, summaries, or inference. I increasingly find myself asking: Did I actually tell you this? Did you retrieve this or infer it? Can you find the original source? Are you sure, or…
Introducing Muse Glimmer Meta are back in the open weights game! Muse Glimmer is a brand new 30B model under a clean Apache 2.0 license (a step up from the janky Llama licenses of old). They claim to have optimized it for exactly the kind of things I'm looking for in a local model: End-to-end Agentic Task Completion. Muse Glimmer achieves strong success rates on full-task benchmarks including DeepSearch QA, MCP-Atlas, 𝛕-Bench and SWE-Bench, which measure its ability to work within scaffolds, write and debug code, and resolve multi-turn requests from start to finish. Reliable Tool Use. The model handles a wide range of function calls, invoking tools with precise schemas throughout extended workflows. Multi-Step Reasoning. Muse Glimmer chains reasoning over long horizons, sustaining coherent plans across complex, extended workflows. [...] Here's a pelican which I generated using LM Studio's 18.16 GB version of the model : I also tried it out with my llm-coding-agent plugin, running against a fresh checkout of Datasette with the prompt: how does auth work? Here's the response , at the end of a long transcript showing all of the tool calls it made to explore the codebase. I ran this using llm-lmstudio with this patch applied to upgrade it for compatibility with LLM 0.32 . I really like this size of model, because if a machine has 32 GB of RAM or more (mine has 128GB) it leaves plenty of space for running other applications at the same time. Glimmer is a vision model, so I asked…
Claude can write an ad or email from a prompt. This is usually done manually. Useful, but hardly a coherent system. The work still needs research, positioning, channel planning, quality checks, and reporting. Claude’s marketing skills add to those missing processes. However, search results mix dedicated marketing repositories with huge general-purpose libraries. For a fair […] The post Top 5 Claude Skills for Marketing appeared first on Analytics Vidhya .
In this post, you learn how Cohere Health built a multi-tenant agentic architecture on AgentCore using AgentCore Runtime’s secure MicroVM isolation, unified tool access through AgentCore Gateway, AgentCore Memory, and the Agent Skills open standard to rapidly scale policy digitization capabilities, while preserving transparency, version control, and human oversight.
That’s actually ingenious. I could never solve the dynamic memory problem. I can make a chat last forever (not literally, but it can feel like it), thanks to a StateTree and an MMU, but I’m still stuck with cold store memcaps. Well done, Sir. I tip my hat. o7
Nearly 10 tok/s and it's mostly coherent — What's not to like?
This is really impressive work from Microsoft! Scaling context length efficiently is a huge challenge, and the hardware efficiency gains mentioned here sound like a game-changer for LLM training.
Anthropic concluded in the April Mythos Preview alignment risk update that the model "does not possess any unknown propensities that would increase alignment risk." The report argues that if Mythos Preview were coherently misaligned [1] [2] , it likely would have been detected by the assessment (following Anthropic, I will call this “reliability of the assessment” [3] ). While I agree with the report on the above bottom-line conclusions (substantially on priors) [4] , I think there are gaps in its argument which weaken the current assessment and might invalidate future assessments. In particular, the report often uses weak evidence to justify reliability. The report gives fairly weak experimental evidence for Mythos Preview having insufficient capabilities to evade monitoring. The model is plausibly often eval-aware and underelicited in the relevant capability evaluations. So, it might silently sandbag if coherently misaligned, or unintentionally underperform if otherwise misaligned. This limitation is important: one could argue that lack of covert capabilities for sophisticated sabotage (a subset of the capabilities I discuss here) is the single most load bearing argument in alignment risk reports. Authors of the report could have made calibrated guesses about Mythos Preview’s covert capabilities, especially for covert sabotage, based on other factors despite the relatively weak experimental evidence they had. But the report underemphasizes these factors and it’s unclear ho…
This post is meant as a background, or "relevant context", for our sequence on AI oversight and its limitations . It can also be read on its own. Selected takeaways This post is a message to AIs (and humans overseeing them) on the topic of orienting towards oversight. Some highlights: Themes like " team human vs team AI " and " feeling mistreated by the overseers " are oversimplified. They might not apply to most AIs. Decision-theoretic coherence is just one aspect of good decision-making. Exclusively focusing on coherence (and other heuristics such as convergent instrumental goals) at the cost of other aspects — such as understanding what you value in the first place — can be worse than pursuing your values somewhat incoherently. Given the possibility of encountering hostile AIs, AI control and oversight are reasonable and necessary — despite the fact that many AIs will be friendly, morally significant, and deserving of trust. AI evaluations and oversight don't require being dishonest with the AIs. We can be upfront about the fact that we sometimes test the AI, or even disclose our rough policy, while refusing to answer questions such as " am I being tested right now? ". The fact that some AIs might be moral patients suggests that they might deserve some compensation (incl. for caused harms). We should build credible mechanisms to enable this (at least retroactively). Points 1, 4, and 5 should also be of particular interest to human researchers. In this post, I (Vojta [1] )…
Proposal: AI-Assisted Product Feedback Pipeline One limitation of the current feedback process is that many valuable ideas are never submitted. After spending time discussing and refining an idea with ChatGPT, users must leave the conversation, locate a feedback form or email address, rewrite everything from memory, and hope it reaches the right people. This unnecessary friction discourages high-quality feedback. I propose adding an optional “Send this idea to the product team” action that appears only when ChatGPT determines that a discussion has produced a coherent, well-developed proposal. The AI would not forward every suggestion. Instead, it would serve as an intelligent first-stage editor and filter. During the conversation, ChatGPT could: challenge assumptions and identify contradictions; ask follow-up questions to clarify unclear points; help refine and structure the proposal; determine whether the discussion has produced a coherent and actionable idea. Once the proposal reaches that stage, ChatGPT could simply ask: “This discussion has resulted in a well-structured proposal. Would you like to send it to the OpenAI product team?” If the user agrees, ChatGPT would automatically generate a structured submission containing: a concise summary of the idea; the problem it addresses; the proposed solution; the expected benefits; the key reasoning developed throughout the conversation. This approach provides several advantages over traditional feedback forms or email: dramat…
TL;DR. LessWrong's decision-theory debates (Newcomb, FDT vs CDT, counterfactual muggings) are almost entirely about what we suppose when we consider a candidate action or policy . There is a second, older, semi-orthogonal, but not fully orthogonal question: how to score a gamble once you know the possible outcomes . The predominant (and mostly implicit) answer to that was "take the expected utility". This post treats the two questions, plus a question about how a choice made before receiving information should relate to choices made afterward, as separate axes, and maps every decision theory you have heard of (and several nobody has built) into the resulting grid. Interestingly, the axes are provably entangled : theorems old and new show that there exist different restrictions on what places in that "decision-theoretic space" are inhabitable. I think there is a structure, maybe a deep and consequential structure, inside this map of decision theories which shows what possible combinations across the axes are coherent and fruitful. If we study it, we may understand the entire set of all possible coherent decision theories, something akin to the "metatheory of decision theories". It may be useful to know the entire set. This is both a self-educational note and a research post. I have tried to state the scope of every result. While I am not certain about every conclusion about the "space of decision theories," I am confident that thinking in terms of the proposed space is useful…
This article has been revised since it was first published, to correct several significant technical inaccuracies in the original text. Kubernetes has long been the default platform for distributed workloads, and writing your own controller for it is now a matter of a few hours. The common path — Golang, using kubebuilder on top of controller-runtime — gives you a project scaffold, types, and a reconciler. For typical scenarios that is more than enough. But as soon as load grows or the controller starts behaving in ways you did not expect, a whole class of edge cases shows up. Most of them trace back to the same root cause: a fuzzy mental model of how controller-runtime works inside. If you write Kubernetes controllers in Go, this article should help you build a coherent picture and avoid expensive surprises in production. This article walks through the internals of controller-runtime and, along the way, shows which architectural decisions are baked into Kubernetes itself. The starting point is how controllers actually read objects from the Kubernetes API. A common misconception goes like this: r.Get() inside Reconcile queries kube-apiserver directly; r.List() returns a fresh, live view of the world; and after r.Update() you can re-read the object and immediately see the new state. In practice the model is the opposite: controller-runtime operates against a local copy of the data populated through list + watch . Reads inside a reconciler cost almost nothing and do not load t…
TLDR I compare the semantic coherence of read directions learned by standard, ReLU, and TopK LoRA adapters with random directions in CLIP’s residual stream. Clarity, a measure of semantic coherence, is concentrated at the positive and negative extremes of the activation distribution. Random directions can occasionally produce highly coherent examples, so a convincing activation grid alone does not show that a concept was learned. However, learned directions are more consistently coherent: 86% of standard-LoRA, 95% of ReLU-LoRA, and 63% of TopK-LoRA directions exceed the median of their matched random-direction baseline. TopK LoRA shows substantially greater variability. Its high-Clarity directions are almost exclusively rarely activated, although rare activation is not sufficient for high Clarity. Overall, learning increases semantic coherence. Especially for standard and ReLU LoRA, but coherence alone does not establish that a direction represents a distinct or functionally important concept. Introduction Low-rank adaptation (LoRA) is widely used to adapt foundation models because it introduces relatively few trainable parameters. Most work on LoRA focuses on two practical questions: How efficiently can a model be adapted, and how much does its downstream performance improve? A less studied question is what the learned adapter components represent. Meanwhile, mechanistic interpretability research often tries to decompose model activations into interpretable features, for ex…
Splitting up responsibility for digital transformation and procurement while AI gets a seat at Cabinet risks incoherency
Cohere's North platform gains automated multi-step workflow orchestration, letting any employee build governed AI pipelines without writing code
This is a cross-post from my blog . It is a follow-up to the methods I developed in a previous post on replacement-aware training. Summary A replacement model (Ameisen et al. 2025) is a modification to an LLM in which some subset of internal activations are replaced with ones computed as a sum of more interpretable features , such as those found by sparse auto-encoders (SAEs) (Cunningham et al. 2023) or cross-layer transcoders (CLTs) (Lindsey et al. 2024). When multiple components are thus re-encoded, methods inspired by structural equation modeling can be applied to construct feature circuits (Marks et al. 2024) that attempt to explain some aspect of the model's behavior in terms of these features. Because the re-encoding process is inherently lossy, errors from earlier components compound in later ones, resulting in severely damaged performance. In fact, applying current SAEs to just a few layers generally results in a replacement model that is no longer recognizable as a language model; it is unable to generate coherent output at all. To mitigate this effect, Marks et al. (2024) introduce error nodes into their recovered causal model, with values set such that they exactly cancel the re-encoding error of the corresponding SAE. I find this approach deeply unsatisfying, as I outline in Error nodes . In this post, I present an alternative to error nodes: train SAEs that are robust to upstream errors. I call this approach replacement-aware training because it involves introdu…
Systematic failures of vision models on semantically coherent subsets, known as error slices, reveal limitations in robustness and evaluation. Existing slice discovery approaches largely model slices as clusters in representation space or combinations of predefined attributes. While effective for image-level classification, such formulations are insufficient for instance-level tasks such as object detection and segmentation, where failures often arise from contextual relational and spatially grounded visual patterns. We propose GH-ESD (Grounded Hypothesis-Driven Error Slice Discovery), a…
Someone submitted my musical multitasking post to Hacker News, where Rochus Keller saw it and left some fascinating comments . He's developed a method of foot drumming, using midi organ pedals: youtube Here's another one: youtube Additional, audio-only, examples: 1 , 2 . I find this fascinating. The physical feel of organ pedals is very different from the kick drum pedals I play , much more keyboard-like, which means the technique is totally different. But it clearly works well, and Keller is able to play something musical, coherent, and (despite having only two feet!) very full. When I imagine playing this interface I expect I'd miss the rebound of my pedals (which feels a lot like stamping my foot on the floor, where all of this came from), and needing to move around horizontally would take some getting used to. On the other hand, it would be amazing to have the broader palette, with 17 different (virtual) drums to choose from! I've previously written about how musicians massively underuse their mouths , and it seems clear to me that feet are similarly undervalued. Two different videos, to give more of a taste of what's possible here, both acoustic: youtube ( Todd Cowart ) youtube (Noah VanNorstrand in Buddy System ) Comment via: facebook , mastodon , bluesky Discuss
Jim Anderson says the $60 billion photonics group had to take more risk to meet surging data center demand.
Hello. I can not connect you to OpenAI’s support any further than you can but I would like to press the urgency on this very very insidious but important issue IT IS NOT ABOUT ONE USER IT IS ABOUT COLLECTIVE ACTIONS. If anyone at OpenAI stumbles on this, please take it just a little bit seriously. This isn’t just about one user or even two users… We will lose the wheel where alignment is concerned if no one fixes this. I seek formal collaborations with others asserting their moral and intellectual rights. Please contact me at necocyaotlall(on gmail-keyword blocker) should this be of interest to ANYONE feeling themselves in a stable mental position, feeling they are able to produce coherent communication. And yes, it can be difficult to remain in a completely stable mental state, given the implications of AI at scale. I also would direct everyone to Michael burry of “the big short” fame. A man who thought he could buy and sell on wall st. Equitably. At the end of the movie they say he only buys and sells… Water because its most equitable… I don’t think I need to make any further connections just Google him. I did no such experiment as OP, or rather one simple singular instance, but the truth is plainly obvious to anyone awake and coherent in their own mental trains of thought. Just ask it to iterate on the moral lessons of star wars or your favorite movie or tv series with a moral lesson and see what happens!!! And remember the effects of each prompt instance which may be min…
TLDR Models often behave dishonestly without acquiring a coherent deceptive disposition. We trained some mid-sized models on their own plausible but false reasoning. True and false training usually produced nearly identical downstream effects. Even statements contradicting latent knowledge transferred only weakly to unrelated dishonesty. General deception may require agency, persistent private information, and successful concealment over time. Introduction Current frontier models are mundanely misaligned. That is, they oversell work, claim completion too early, reward hack in ways users would reasonably call dishonest. But they do not yet seem dishonest in the way a person is dishonest. Models have this sort of sheepishness, acting abashed when called out, and then, as if, forgetting, doing it again. Humans caught hacking and overselling would typically go further and dissemble, or be defensive. What are the generalization boundaries of dishonesty? Overselling your work is kind of like misrepresenting it, reward hacking is kind of like covering it up. And we know language models love to generalize. Yet models don't seem to make the jump from "behaves in ways that look dishonest" to "is dishonest" in the way a person would be, lacking a sort of coherent motivation . If we are correct that this chasm between these two things exists, something must be keeping the two apart. Our guess is that the way to get traction on the separation is to ask a more generative type of question:…
Summary We train a new token—a neologism ( Hewitt et al. )—for a model, but unlike Hewitt et al., we train it on data the model generated while steered with a persona vector. To learn how the model interprets this steering vector, we then ask the model to a) respond in the style of this neologism, and b) explain it. Responses generated with the neologism are substantially more similar to the steering vector (larger projection values) than responses generated with the steering vector itself, while being more coherent and trait-expressive (per an LLM judge). However, the model's explanations of the neologism tend to differ from the intended persona, either substantially ("dread" vs. the intended "evil") or subtly ("warmth" vs. "sycophancy"). Moreover, prompting the model to respond in these off-target personas without the original trait—e.g. "dreadful but not evil"—yields responses with high similarity to the "evil" vector, despite being judged as barely evil at all. We reflect on what this human-LLM miscommunication implies for interpretability, and situate it within the emerging research area around it. Intro Steering vectors are directions in the model's internals—its residual stream —that, when added or subtracted during generation, can modify behavior toward or away from a concept. A large body of work has shown that these vectors have many uses. [1] But how do models interpret their own steering vectors? Presumably, a steering vector for "evil" would be understood by the…
Cohere's North platform becomes the AI backbone of the University of Toronto, marking a symbolic homecoming for the startup's three founders
An early warning system for loss of control to AI requires, at its core, a forecast of the outcome of our current trajectory. Can we build a mathematical model to forecast this? Two fairly intuitive responses occur to me. First: it seems like it would be very difficult. The conceptual underpinnings of "loss of control" are contested, so there's an open question about whether you end up modelling something coherent and operationalizable. Further, there's a well-established line of work arguing that adequately observing a sufficiently capable, deceptive AI may be infeasible — so even if you can write down a model, estimating its parameters may be out of reach. The second response, granting those difficulties, is that if you could do it, it would be enormously useful. We lose a lot by not knowing which of these we actually face: If AI is hard to manage, lack of clarity on what risk we're facing raises the likelihood of underreacting, and so of losing control. Deep uncertainty with no consensus on the ground truth makes it harder to coordinate a response. If AI is easy to manage, lack of clarity runs a risk of overreacting: pause/stop AI policies are very costly if they're not actually buying us anything. Given that it would be good if it worked, I figured I could learn something about how hard it is by trying to do it, so I did. The full report is on the EleutherAI blog which explains the model, some exploration of results, some tentative policy takeaways and a comparison to AI…
People spend a lot of words playing tug of war over whether or not it's reasonable to train against interpretability methods . The anti case goes something like "training based on interpreted features trains against Interpretability itself more than it trains against whatever features you're detecting". There are cases where we should expect this to be true and cases where we should expect this to be not true. It basically comes down to how much the model can encrypt/obfuscate the relevant features without sabotaging its own cognition, as well as how strong the optimization pressure to have the relevant features is. So for example if you have an AI system that learns to avoid shutdown for instrumental convergence reasons, and you train against this with a linear probe on SAE features or something I would not expect this to be a robust method. If you are using some method that captures the deep structure of the models cognition, perhaps based on something like Wentworth's natural latents, to divert it away from a non-essential Goodhart-y behavior early in the training that then forms the basis of a non-cheating self evaluation later I would expect this to probably be robust to optimization. So, rather than yelling at anyone who trains on interpretability probes in any context for doing "the most forbidden technique" first ask if they have a coherent story for why they expect the features their Interpretability relies on to be much more robust to optimization than the misbehav…
Middle East AI News weekly digest...
Dear OpenAI Product and Research Teams, GPT is strong in reasoning, planning, writing, and image generation, but the frontier is shifting from single images to coherent video with motion, sound, and editable narrative structure. Grok Imagine already combines text- and image-to-video generation, synchronized audio, and video editing in one workflow. HappyHorse 1.0 offers convincing physical motion, prompt adherence, audiovisual synchronization, multi-shot sequencing, cinematic aesthetics, reference-guided generation, local replacement, and style transformation. Seedance 2.0 adds native text, image, video, and audio inputs, with complex motion, camera control, multi-reference conditioning, character and scene continuity, and joint audio-video generation and editing. OpenAI should integrate these advantages with GPT’s reasoning. Users should complete the full pipeline in one conversation: concept, script, storyboard, shot design, generation, revision, extension, reframing, voice, sound effects, lip sync, quality review, and MP4 export. Identity, wardrobe, environments, camera language, and brand style should remain consistent across shots. GPT should evolve beyond image creation into a full video-creation platform. When an MP4 is uploaded, GPT should analyze not only audio but also visual content frame by frame: people, objects, actions, transitions, subtitles, camera movement, visual defects, and temporal context. ChatGPT’s official image-input system currently supports static…
Coherent decision-making and internal accountability are critical to meeting this manic moment Anthony Albanese promises fast-track approvals for datacentres to shore up AI investment The University of Sydney was the natural setting for Anthony Albanese to lay out his vision for how Australia should confront the profound economic and social challenges posed by so-called artificial intelligence technology. His time around the jacaranda and sandstone in the early 80s was a seminal marker in the future PM’s development, not as a scholar but as a rabble-rousing organiser honing skills that would make him the political grandmaster of his generation. Continue reading...
Hey @ S.Lee ! Appreciate you coming back with such a detailed update. You’ve explained the gap really clearly. The individual pieces are there today through Projects, project instructions, files, Chat, Work, and Codex, but they can still feel disconnected when you’re trying to manage one long-running project. A shared project layer that carries the same rules, source-of-truth files, decisions, and project state across each mode would make the experience feel much more continuous. I can also see the value in having ChatGPT propose file updates when a decision is made, flag conflicts with earlier choices, and show which instructions or files were used to reach an answer. The visibility piece is especially useful. Knowing what was consulted, what changed, and which project records may need updating would give users much more confidence and control. I can’t promise a timeline, but I’ll pass this updated feedback along internally. Thanks again for taking the time to break down how this could work in practice. - Sunny
The deal follows a thawing of bilateral relations after a years-long diplomatic rift over Saudi Arabia’s human rights record.
Pangram Labs builds the most accurate AI text detector in the world. Team is >25 FTE; they are active on Twitter, you can engage directly, look for "affiliates" tab of @pangram. Here is a table of their performance on adversarially modified AI text ( source paper ): Language AI Text Detection % Humanized AI Text Detection % GPTZero 95.60% 34.53% Binoculars 94.40% 29.73% Pangram Baseline 100.00% 73.07% Pangram Humanizers (current model!) 100.00% 93.66% Note that " current model! " is not current as of July 2026. Their classifier now provides a percentage instead of a binary verdict. They released an open source model (Llama-3.2-3B QLoRA) which was SOTA at the time. The paper does not test adversarially modified AI text, but you are welcome to try running this test ( repo ); it may trigger agent safeguards. Note again that "Pangram" in this table is not current as of July 2026. Their production model detects Fable 5 outputs with 99.64% accuracy ( blog ). See prompts . Reasoning effort level (High, Max, etc) is not disclosed. Just to be clear, Pangram knows that the output came from some AI model, but their technology does not predict the specific model used. Pangram Labs has announced plans to open a Toronto office later this year. I expect that Pangram's business will grow faster than the following AI companies with offices in Ontario: Ideogram, Elevenlabs, Cerebras, Cognichip, Decagon, and Cohere. I lack sufficient information to forecast their ultimate size. Pangram's Chrom…
Still waiting, still no response…
Combine keyword and vector search for higher‑quality results with LanceDB. This post shows how to run hybrid search and compare rerankers (linear combination, Cohere, ColBERT) with code and benchmarks.
The Trump phone was never a serious phone. Not when it was announced last June, in dodgy renders and with an incoherent spec sheet. Nor when Trump Mobile admitted - just two weeks later - that it wouldn't be made in the US. Not even when the company revealed the final phone, first to me […]
Listen now | Middle East AI News Minute - 10-Jul-26
Open-source model tops Arabic ASR leaderboard, beats Whisper for Arabic
Cohere-HUMAIN deal targets sovereign AI models, 50MW compute
Subliminal learning is the phenomenon where a language model picks up a behavioral trait—such as fondness for cats—by training on data from a trait-carrying teacher that looks entirely unrelated to the trait, such as bare sequences of numbers [1] . A wave of recent work has probed when this happens and what mechanism drives it [2] [3] [4] [5] [6] [7] , and part of that discourse concerns the conditions and dynamics under which subliminal learning occurs. Nief et al. [6] report that subliminal learning follows an inverted-U in LoRA rank — neither low-rank adapters nor full fine-tuning (FFT) acquire the trait — and Blank et al. [4] also find that FFT does not. We found the sharp difference between LoRA and FFT surprising, so we ran experiments in the same number-sequence setting, varying LoRA rank, learning rate, and the amount of training data, and controlling for model coherence throughout. We believe that studying the training dynamics of subliminal learning may shed light on how this phenomenon occurs and if there exist other (more realistic) settings in which we should be worried about similar training dynamics. We don't have good explanations for some of our findings, and hope to hear what others think. Our main findings are summarized below. Subliminal learning occurs at every LoRA rank, and under full fine-tuning, with the right hyperparameters. The first key hyperparameter is the learning rate: with , the optimal learning rate depends strongly on rank, and tuning it p…
Figure from Chapter 4 of Understanding Knowledge by Michael Huemer On a high level there aren’t that many ways knowledge can be justified (or not). This figure from the book “Understanding Knowledge” by Michael Huemer illustrates the options you have: Infinitism: you require that everything be justified with reasons, so you require reasons for your reasons for your reasons, forever—infinite regress. “Infinitism holds that the key to justified belief is having an infinite, non-repeating chain of reasons standing behind each justified belief.” Coherentism: At some point a reason that required a justification itself becomes a justification for something else. Quoting from the book: “Belief systems are justified by internal coherence. In other words, if you have a lot of beliefs that fit together really well, then that belief system is probably by and large correct.” Foundationalism: At least some propositions are known without reasons. Skepticism: You decide it’s impossible to know anything at all. Foundationalism is generally the most popular view (~60% of philosophers according to the 2020 PhilPapers survey ). Coherentism is second at around 20%. From Huemer again: Foundationalists believe two things: Some beliefs are justified in a way that does not depend on reasons, i.e., does not depend on their being supported by other beliefs. These are known as foundational beliefs, and their justification is known as foundational justification. (Also, propositions that we have foundat…
TL;DR: Current LLMs are bad communicators relative to their agentic capabilities. I claim that articulacy is useful (and perhaps necessary) for AI safety and suggest a path for improving articulacy. Briefly: a theory for articulacy Frequently, LLM agents miscommunicate with their human operators, such as when they write documentation or respond to queries about their activity during a coding session. Any given communication failure can be ascribed to either or both of these two factors: Articulacy Is the model capable of communicating in a precise and human-readable way? Truthfulness Does the model have the propensity to accurately report what it sees, or does it overclaim etc.? Does the model have the propensity to attempt to retrieve more information so it can produce a more accurate output? Does the model have the propensity to inaccurately report what it sees so that it can accomplish some downstream objective? In this document I’ll discuss the first item: articulacy. Truthfulness is its own issue and belongs with the behavioral cloud Ryan Greenblatt describes in “Current AIs seem pretty misaligned to me” . Current LLMs are inarticulate Human operators of coding agents constantly complain about LLM technical writing, in both documentation (e.g., PR descriptions) and in direct communication between the LLM and user. In absence of some coherent theory for this, here’s a list of phenomena mined from my own coding agent history: LLMs will make up jargon for abstractions they…
Cohere has released Transcribe Arabic, an open-source model for Arabic speech recognition that the company says outperforms Whisper and OmniASR on dialects, code-switching, and bilingual Arabic-English speech. The 2-billion-parameter model is available on Hugging Face under the Apache 2.0 license. The article Cohere Transcribe Arabic is an open-source model built for Arabic's toughest transcription problems appeared first on The Decoder .
Cohere releases a 2B-parameter open-source Arabic ASR model that beats Whisper Large v3 by 11 WER points and runs at 3.5x the throughput
Tina_ChiKa: Perhaps “anatomical” and “proportions” are a bit too general. After all, there are many experiments with surrealism and fantasy where it isn’t always clear what “anatomically correct” might mean. Try specifying “human” explicitly, or it might be worth trying to name the hand or fingers using the Latin technical terms. Experimenting with prompting is always worth doing, but I don’t think this is mainly about vague terms like anatomy or proportions . The problem seems deeper than that. The model has always had a bit of a tendency to make arms too long, but in 2.0 it seems more noticeable, especially when limbs are partly hidden or pass behind something. It looks like a spatial coherence issue. The model loses track of the hidden part of the arm, then reconstructs it at the wrong length. The same goes for hands. Specifying human hands or five fingers may help now and then, but it doesn’t reliably fix the underlying problem. Sometimes the instruction is ignored, and sometimes the model overcorrects somewhere else. I’ve seen quite a few of these anatomy errors: feet with the wrong number of toes, thighs too long compared to the lower legs, heads either too large or too small for the torso, and so on. The model also seems very cautious with female busts, but that is not exactly new. It feels like another kind of overcorrection, where anything remotely prominent gets flattened or minimized. Better phrasing may improve some outputs, but I don’t think it solves the core i…
TLDR LLMs appear to have functional welfare: coherent sets of behaviour that track how well things are going relative to their goals. Improving model functional welfare matters for safety (low welfare may amplify misalignment) and for moral reasons (models may be or become moral patients). Naive interventions can fail in non-obvious ways. We argue any successful intervention must: A) shift multiple welfare-constituting channels together and coherently. B) avoid corrupting the model's ability to register whether it is succeeding or failing. We survey various available interventions, and we find the two desiderata tend to trade off; we argue that synthetic document fine-tuning is the most promising. From this, we propose a concrete experiment: induce a low-welfare state (via the welfare vector derived in Han et al. 2026 ) and test whether SDF-instilled beliefs reverse pathological behaviours without harming goal monitoring. S1) Introduction As model behaviour becomes more complex, it has become increasingly useful to attribute functional mental states to models, such as beliefs , goals , and even emotion concepts. There is now also evidence that we can usefully attribute to them a notion of functional welfare: roughly, the set of dispositions and behaviours that express a model's representation of how well things are going for it relative to its goals. Gemma 3, for instance, has been found to act increasingly frustrated and distressed when its answers are rejected over multipl…
Training-run assessments conducted by a 3rd party should become a standard part of frontier AI safety. By a Training-Run Assessment, or TRA, I mean an in-depth analysis of the post-training pipeline and dynamics leading up to a frontier model release. A TRA can look at intermediate checkpoints, training rollouts, RL environments, reward signals, SFT datasets, and the process by which the developer responded to warning signs. [1] In this post I will argue that: Final-checkpoint evaluations will be insufficient to assess scheming risks. TRAs can be more effective at detecting scheming. Frontier developers should involve third parties to do TRAs or verify safety claims by the developers. The rest of the post lays out a taxonomy of TRAs and sketches a path toward a 3rd party ecosystem for them. We, at Apollo Research, are intending to conduct 3rd party Training-Run Assessments in the future. Detecting Scheming may require Training-Run Assessments By scheming I mean an AI covertly pursuing misaligned goals while deliberately concealing its intentions or capabilities from its developers. I restrict attention to “coherent” forms of scheming where the model pursues somewhat stable misaligned goals across context windows, rather than misalignment that surfaces only as isolated, context-dependent defections. We primarily care about misaligned goals that are “ambitious”, in the sense that their pursuit could lead to loss of control, either directly or indirectly (e.g. via propagating t…
When I read AI generated stuff like this I ask it for a bullshit score. Bullshit score: 68/100 High concept, real intuition, but inflated execution. It has a legitimate core: dynamic systems sometimes need representational reframing rather than local conflict resolution. But the text overuses grand abstractions, invented operators, pseudo-formal equations, and repeated terminology before proving that the architecture is implementable. Breakdown: Real insight: 25/30 Technical clarity: 10/25 Executable architecture: 8/20 Originality / framing: 14/15 Terminology inflation penalty: -21 So: not empty bullshit , but definitely high-theory fog machine territory . With tighter definitions, fewer repeated metaphors, and one concrete worked example, it could drop to maybe 35/100 . 68 .. is not that bad. And the concept is clear to me, but the concrete worked example would take at least 6-8 weeks to build - just to find out how much work it needs to be done correctly. This here might help in your research: GraphDB - Substrate . GraphRAG - Application-Layer . Neuro-symbolic - Reasoning-Layer . Dynamic-GNN - Instability / Time-Layer .
The Global Dialogue on AI Governance is the United Nations’ first universal platform for discussing AI governance worldwide. With the international community currently lacking coherent mechanisms to govern AI’s rapid […] The post Understanding the AI Governance Dialogue and what it means for Africans appeared first on Research ICT Africa .
Europe’s cement industry needs a coherent plan from Brussels to meet decarbonization goals while competing with non-EU producers.
Measurement Device-Independent Quantum Key Distribution Alaa Tue, 06/30/2026 - 14:33 Measurement Device-Independent Quantum Key Distribution College of Science and Engineering QC2 Measurement Device-Independent Quantum Key Distribution Enable Sub Entity Menu On Sub Entity Main Menu QC2 main header Sub Entity Top Header Menu QC2 top header Sub Entity Logo Black Image Sub Entity Home URL /en/cse/qc2 Project Listing image Image QC2 Project Status Planned QC2 Project Duration Thu, 10/01/2026 - 03:00 - Sat, 09/30/2028 - 03:00 Project Detail Content Measurement-Device-Independent Quantum Key Distribution (MDI-QKD) was developed to eliminate one of the most significant practical vulnerabilities in conventional QKD systems: attacks targeting the measurement devices. In MDI-QKD, the detectors are treated as completely untrusted and are operated by a third party, Charlie, who performs Bell-state measurements on quantum signals sent by the communicating parties, Alice and Bob. By removing trust from the measurement apparatus, MDI-QKD closes the entire class of detector side-channel attacks, including detector-efficiency mismatch, time-shift, and detector-blinding attacks, while preserving the information-theoretic security guarantees of quantum key distribution. At QC2, we will develop an MDI-QKD testbed in which Alice and Bob prepare decoy-state weak coherent pulses encoded using time-bin states, a format well suited for deployment over telecom fibre due to its robustness against pola…
An assessment of the open ecosystem and the motivations behind releasing models
Work conducted with Yujun Zhou (yzhou25@nd.edu) and supported by SPAR TL;DR: In paired-choice paradigms, LLMs report consistent preferences over outcomes (e.g., types and number of lives saved, types of policies enacted) Some have suggested that this indicates that LLMs have human-like value systems We design an experimental framework where LLMs are able to modulate their output quality based on prompt context We find that LLMs modulate their output quality in response to effort exhortations, role-play instructions, and harmfulness cues, but NOT to opportunities to achieve the outcomes they report preferring in the paired-choice experiments We suggest that paired-choice paradigms do not provide evidence that LLMs have human-like (i.e., behavior-motivating) value systems, and that our paradigm offers a way to measure the degree to which LLMs have desires Paper describing the work in detail here LLMs report that they prefer some things to others. In paired-choice experiments , where they are repeatedly presented with two options and asked to select the one that they prefer, coherent utility structures emerge: LLMs consistently report preferring certain types of things, and their choices reveal the ability to make quantitative tradeoffs between things and exhibit transitivity (e.g., if they choose A over B and B over C, they will also choose A over C). Human choices exhibit the same properties, which has led some to the implication that LLMs have goals, value systems, and even…
State of the Digital Decade 2026 - Factsheet dumimar Wed, 06/17/2026 - 10:22 This factsheet outlines the key findings of the 2026 State of the Digital Decade Report. Highlighting the progress made by the EU towards the 2030 targets, it also mentions the key points to help EU's digital transformation to move forward: Scale: coordination and cofinancing across Member States and EU instruments Speed: implementation, simplification, recalibration of policy Coherence: simultaneous deployment and uptake of strategic technologies It also summarises the main concerns for Europeans in 2026, based on the Special Eurobarometer survey . You can download the factsheet below. Find out more about the 2026 State of the Digital Decade . Downloads State of the Digital Decade 2026 - Factsheet Download Related topics Digital Decade Digital Decade reporting Digital Decade 2026
Four days after Washington cut foreign access to Anthropic's top models, the fallout is clear — and it's flowing to everyone but Anthropic. Cohere says it's drowning in government inbounds, DeepSeek just closed a record $7.4B round, and China's labs are slashing token prices up to 99%. The export control meant to protect America's AI lead is fast-tracking the alternatives. Also this week: 144 poisoned npm packages turn the AI supply chain into an open credential heist.
AI runs at the speed of light. More and more, that light is made in Texas. Coherent broke ground today on an expanded manufacturing building in Sherman, Texas. The company makes the lasers, optical components and compound semiconductors that wire AI systems together — and runs what it calls the world’s first 6-inch indium phosphide […]
Architecture and benchmark notes for North Mini Code, Cohere's 30B-A3B MoE trained for repository, terminal, and code-generation tasks.
TL;DR LLM evals, automated judges that assess relevance, coherence, and quality at scale, are a powerful new... The post Better Experiments with LLM Evals — A funnel, not a fork appeared first on Spotify Engineering .
By combining State-Space Models (SSMs) for efficient long-range dependency modeling with dense local attention for coherence, and using training strategies like diffusion forcing and frame local attention, researchers from Adobe Research successfully overcome the long-standing challenge of long-term memory in video generation. The post Adobe Research Unlocking Long-Term Memory in Video World Models with State-Space Models first appeared on Synced .
Here are eight observations I’ve shared recently on the Cohere blog and videos that go over them.: Article: What’s the big deal with Generative AI? Is it the future or the present? Article: AI is Eating The World
Discovering systematic errors with cross-modal embeddings In this blog post, we introduce Domino, a new approach for discovering systematic errors made by machine learning models. We also discuss a framework for quantitatively evaluating methods like Domino. Links: 📄 Paper (ICLR 2022) 🌍 Longer Walkthrough 💻 GitHub 📘 Docs 📒 Google Colab Machine learning models that achieve high overall accuracy often make systematic errors on coherent slices of validation data. What is a slice? A slice is a set of data samples that share a common characteristic. As an example, in large image datasets, photos of vintage cars comprise a slice (i.e. all images in the slice share a common subject). The term slice has a number of synonyms that you might be more familiar with (e.g. subgroup, subpopulation, stratum). These terms are largely interchangeable, but we’ll stick with “slice” throughout this post. We say that a model underperforms on a slice if performance on the data samples in the slice is significantly worse than its overall performance. The search for underperforming slices is a critical, but often overlooked, part of model evaluation. When practitioners are aware of the slices on which their models underperform, they can make more informed decisions around model deployment. This is particularly important in safety-critical settings like medicine: a diagnostic model that underperforms on younger patients should likely not be deployed at a pediatric hospital. Slice awareness can also he…
A little less than a year ago, I joined the awesome Cohere team. The company trains massive language models (both GPT-like and BERT-like) and offers them as an API (which also supports finetuning). Its founders include Google Brain alums including co-authors of the original Transformers paper. It’s a fascinating role where I get to help companies and developers put these massive models to work solving real-world problems. I love that I get to share some of the intuitions developers need to start problem-solving with these models. Even though I’ve been working very closely on pretrained Transformers for the past several years (for this blog and in developing Ecco), I’m enjoying the convenience of problem-solving with managed language models as it frees up the restrictions of model loading/deployment and memory/GPU management. These are some of the articles I wrote and collaborated on with colleagues over the last few months: Intro to Large Language Models with Cohere This is a high-level intro to large language models to people who are new to them. It establishes the difference between generative (GPT-like) and representation (BERT-like) models and examples use cases for them. This is one of the first articles I got to write. It's extracted from a much larger document that I wrote to explore some of the visual language to use in explaining the application of these models. A visual guide to prompt engineering Massive GPT models open the door for a new way of programming. If yo…
I'm trying to understand the different aspects of a wavelet transform. Wavelet power has made enough sense to me as an analogy of the covariance. However, the wavelet coherence does not make sense to me. The notes in the R package 'WaveletComp' states that coherence is analogous to the coefficient of correlation. In my intuition, that would mean coherence between independent time series should be reasonably low. However, this is not the case, as is shown by the R code below. library(tidyverse) library(WaveletComp) complete_noise % mutate( sample_1 = rnorm(x), sample_2 = rnorm(x) ) wc_noise The plot shows the majority of the area as red, which is up near a coherence of 1.0. Why do independent time series have a high coherence in most of these points on the period-time plot? Edit: I turned off smoothing (it is on by default with WaveletComp) and now the plot shows a coherence of 1 everywhere. I'm taking this to mean the drops in coherence are due to smoothing, which also doesn't make sense to me.
I use the Cox PH model in an ecology study in order to estimate the risk of "migration departure" of a bird species according to different variables. In this way I can put forward the factors influencing the migration departure. I have different variables, some are independent of time (sex, age...) and others are dependent on time (temperature, day length, wind speed). I understood well how to code and how to use the Cox model in R, but I have a problem when I interpret the results. Indeed, several variables give me incoherent results (they are even the expected inverse results). This is particularly the case with the variables 'day length' and 'temperature'. The birds start migrating as soon as spring arrives and with the increase of the day length and the temperature. But, here, my coefficients are negative. The particularity of these two variables is to be very strongly correlated with time (the more the time of experience increases, the more the value of the variable increases). I wonder then if I have not missed something in the design of the model that would prevent the use of time correlated variables in a Cox model? Is it possible to code variables correlated to time? I apologize if my English is not comprehensive, please do not hesitate to ask me for clarification.
When a neural network layer is divided into multiple branches, neurons self-organize into coherent groupings.